diff --git a/paddle/fluid/operators/reader/buffered_reader.cc b/paddle/fluid/operators/reader/buffered_reader.cc
index d5a7c50d95cb82009e8055f0279d858f690059a5..971db8b37d8409185a4e81d1e77b0fc53534e9f5 100644
--- a/paddle/fluid/operators/reader/buffered_reader.cc
+++ b/paddle/fluid/operators/reader/buffered_reader.cc
@@ -85,6 +85,10 @@ void BufferedReader::ReadAsync(size_t i) {
           memory::Copy(boost::get<platform::CUDAPlace>(place_), gpu_ptr,
                        boost::get<platform::CUDAPinnedPlace>(cpu_place),
                        cpu_ptr, size, stream);
+        else if ((platform::is_gpu_place(cpu_place)))
+          memory::Copy(boost::get<platform::CUDAPlace>(place_), gpu_ptr,
+                       boost::get<platform::CUDAPlace>(cpu_place), cpu_ptr,
+                       size, stream);
         else
           // if cpu place is not pinned, async copy is slower than sync copy,
           // so we use sync copy instead.